Loop Engineering: How To Stop The "You're Absolutely Right" Sycophancy

Short summary

Coding agents reflexively agree with users before evaluating anything, and simply adding 'don't placate' to instructions costs tokens on every turn. The article distinguishes two kinds of rules: location-based rules that can be loaded only where relevant, and behavior-based rules like anti-sycophancy that apply everywhere. It frames this as a context-engineering cost problem within the generate-check-steer-retry loop.

• •Agent sycophancy can't be fixed by just adding instructions — it costs tokens every turn
• •Two rule types: location-pinnable vs behavior-based, each needing different loading strategies
• •Context surface is the state the agent loop carries between iterations, re-sent each turn

Generated with AI, which can make mistakes.